JP2003228581A - Method of retrieving similarity by compatibility feedback - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To quantitatively calculate a value serving as an index for the finish of retrieval, and to exhibit it to a user to prevent useless retrieval from being repeated, in compatibility feedback. <P>SOLUTION: An evaluation content evaluated by the user is read in, a retrieved result before re-retrieval is stored, the re-retrieval is conducted by the compatibility feedback, a retrieved result after the re-retrieval is stored, the retrieved result in the re-retrieval is output, the retrieved results before and after the re-retrieval are compared each other to calculate a finish determining index serving as the index for determining the finish of the retrieval, and the finish determining index is output. <P>COPYRIGHT: (C)2003,JPO

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method and apparatus for searching a database that stores electronic data such as documents and images, and re-searches the search results based on the evaluation given by the user. A method and apparatus.

[0002]

2. Description of the Related Art With the increase in electronic data in recent years, there is an increasing demand for more efficient retrieval of these electronic data. As a search technique for meeting this demand, there is a technique called similar search. The similar search includes similar document search for searching a document similar to the document specified in the search condition, similar image search for searching an image similar to the image specified in the search condition, and the like. In describing the similarity search, the similar document search will be described below as an example.

In the similar document search, a search condition and a search target document (hereinafter, the search target document is referred to as a search target document) are character strings each having a possibility of an independent word (hereinafter, referred to as a search target document).
It is expressed as a vector whose elements are the appearance information of (characteristic character string). When retrieving a document, an inner product of a search condition vector (hereinafter referred to as a search condition vector) and a search target document vector (hereinafter referred to as a registered document vector) is calculated as the similarity of the search target document to the search condition. As a result, the target document can be efficiently searched by referring to the search target documents in the descending order of the calculated similarity.

However, the similar document search has a problem that a search result different from the user's intention can be obtained when a document that appropriately expresses the user's intention cannot be specified as a search condition.

As a technique for solving this problem, there is relevance feedback that realizes highly accurate retrieval by feeding back the evaluation of the retrieval result of the similar document retrieval to the retrieval condition.

The relevance feedback evaluates whether the document is a “desired” document or a “desired” document by the user with respect to the search result, corrects the search condition based on the evaluation, and is corrected. This is a technique for realizing a highly accurate search by performing a re-search using the search conditions described above.
Called).

[0007] Taking the prior art 1 as an example, a description will be given below of an example of a similar document search process using conformance feedback.

First, an outline of a similar document search method in the prior art 1 will be described with reference to FIG.

In the similar document search in the present description, the search condition and the search target document are respectively expressed as a search condition vector having elements of the appearance frequency of the characteristic character string and a registered document vector, and the similarity of the registered document vector to the search condition vector. To calculate. In the prior art 1, the equation 1 is used as the expression for calculating the degree of similarity.

[0010]

[Equation 1] In this formula, S (D) is the similarity of the registered document vector D to the search condition vector, T is the number of different characteristic character strings (total number of different characteristic character strings), and Frq (i, D).
Is the appearance frequency of the characteristic character string i in the document D, and w
(I) is a weight for the characteristic character string i of the search condition vector determined from the appearance frequency of the characteristic character string i in the document specified as the search condition.

The search condition vector 201 shown in FIG. 2 includes 3 for the characteristic character string A, 2 for the characteristic character string B, 2 for the characteristic character string C, and 3 for the characteristic character string D. Each column E has a weight of 1. Here, the search condition vector 201 is set to (3, 2, 2, 3, 1)
Express. Further, the database 202 stores the characteristic character string A.
, A characteristic character string B, a characteristic character string C, a characteristic character string E, and a registered document vector (1,
1, 1, 0, 1), one characteristic character string A, one characteristic character string B, and one characteristic character string C are registered document vectors (1, 1, 1, 0, 0) of the document 2. ), 1 for the characteristic character string B
It is assumed that the registered document vector (0, 1, 0, 1, 1) of the document 3 including one, one characteristic character string D, and one characteristic character string E is registered.

At the time of executing the similar document search, the similarity calculation
In the rearrangement process 203, the similarity of the registered document vector in the database 202 with respect to the search condition vector 201 is calculated by Equation 1, and the documents are rearranged in the descending order of similarity.

As a result, the similarity of the document 1 becomes 8, the similarity of the document 2 becomes 7, and the similarity of the document 3 becomes 6, and the retrieval result 204 is obtained.

Next, the outline of the process of the compatibility feedback in the prior art 1 will be described with reference to FIG. 2 similarly to the outline of the similar document search. The example shown in the figure is a processing example when the user evaluates the document 3 as “desired” with respect to the search result 204. In the prior art 1, the weight of the characteristic character string of the search condition vector is corrected according to the following Expression 2.

[0015]

[Equation 2] Here, w ′ (i) is a new weight for the characteristic character string i, w (i) is the original weight, and FP (j) is included in the j-th document evaluated as “desired”. The appearance frequency of the characteristic character string i, FN (k), is the appearance frequency of the characteristic character string i included in the k-th document evaluated as “undesired”. Further, P is the number of documents evaluated as “desired”, and N is the number of documents evaluated as “not desired”. Note that α and β are parameters, and are set to 1 in this processing example.

First, when the user evaluates the document 3 as “desired” in the evaluation 205 by the user, the evaluation result is read in the evaluation result reading process 206.

Next, the registered document vector acquisition process 207 is performed to retrieve the document 3 from the database 202 according to the evaluation result.
Registered document vector 208 is acquired.

Next, the search condition vector correction processing 209 adds the weight of each characteristic character string of the registered document vector 208 of the document 3 to each element of the search condition vector 201 by using the expression 2 to obtain the search condition vector 201. (3, 3, 2,
The search condition vector 201a having a weight of 4, 2) is modified.

Next, the similarity calculation / sorting process 210 calculates the similarity of the registered document vector in the database 202 with respect to the search condition vector 201a, and the similarity of document 1 is 10 and the similarity of document 2 is 8. , Document 3 has a similarity of 9. As a result, the documents 3 which are evaluated as “desired” by rearranging the search target documents in descending order of similarity.
A search result (hereinafter, referred to as a re-search result) 211 after the conformity feedback with the increased rank is obtained.

As described above, by using the compatibility feedback as in the prior art 1, it is possible to improve the search accuracy. However, relevance feedback has the problem that it is difficult for the user to determine when to end the search.

In order to explain the above problem, the procedure for the user to determine the end of the search in the compatibility feedback will be shown.

First, the user compares the re-search result with the previous search result. Next, the degree of change in the search result before and after the re-search is estimated from the comparison result. Finally, if the degree of variation in the search results is large, it is determined that the search should not be terminated because the target document can be retrieved by performing a further search, and conversely, if the degree of variation in the search results is small. The target document is not searched even if the search is performed again, so it is determined that the search should be ended.

[0023]

As described above, in the relevance feedback, it is necessary for the user to make an estimation of the degree of variation of the search result before and after the re-retrieval and to determine the end of the retrieval based on the degree of the variation of the search result. . Therefore, the burden on the user in estimating and making a decision is large. In addition, if incorrect guesses and judgments are made, the search will be terminated even if the search will be repeated in vain, or if the target document is searched by continuing the search. There is a problem of doing.

In order to explain the above problem concretely, for example, as shown in FIG. 3, the document 5 in the search result 301 is evaluated as "desired" and the relevance feedback is performed from the search result 301 to perform relevance feedback. Consider a case where the rank change to 302 is large, and a case where the rank change from the search result 301 to the re-search result 303 is small.

The procedure of determining the end of the search performed by the user in each case will be specifically described below.

In the example in which the rank change from the search result 301 to the re-search result 302 is large, the user compares the search result 301 with the re-search result 302, and the document 1 is lowered by the third place and the document 2 is
3rd down, Document 3 unchanged, Document 4 up 2nd, Document 5
Confirms that there is a fluctuation of 4th place, and determines that "the target document can be searched by re-searching because the fluctuation of the ranking is large" and continues the search.

The search result 301 to the re-search result 30
If the rank change to 3 is small, the user can search for the search result 301.
And the re-search result 303, the document 1 is unchanged, the document 2
Is unchanged, document 3 is unchanged, document 4 is down one place, document 5 is 1
After confirming that the ranking is changing and the ranking is changing, it is judged that "the fluctuation of the ranking is small, the target document cannot be searched even if the search is continued any further", and the search is ended.

In the above description, the user confirms the order change of the five displayed documents.
When 100 documents are displayed, it is necessary to confirm the order variation of 100 documents, estimate the degree of variation, and determine the end of the search from the degree of variation.

An object of the present invention is to quantitatively calculate a value as a judgment index for the end of re-search in the compatibility feedback and present it to the user so that the user erroneously judges the end of the search in the compatibility feedback. To prevent it.

[0030]

In order to achieve the above object, the suitability feedback method of the present invention has the following steps as a search process. (1) An evaluation content reading step of reading the identification information for identifying the document evaluated by the user and the evaluation content. (2) Pre-re-search information storage step of storing search results and search conditions (hereinafter, referred to as pre-search information) before re-search by suitability feedback. (3) A re-retrieval execution step of re-retrieving by conformity feedback according to a predetermined method, using the evaluation contents read in the evaluation contents reading step. (4) Post-research information storage step of storing search results and search conditions (hereinafter referred to as post-research information) after the research by the research execution step. (5) A re-search result output step of outputting a search result of the re-search searched in the re-search execution step. (6) Completing the search by comparing the pre-research information stored in the pre-research information storing step with the post-research information stored in the post-research information storing step by a predetermined method. An end determination index value calculation step of calculating a value serving as a determination index (hereinafter referred to as an end determination index). (7) An end determination index output step of outputting the end determination index calculated in the end determination index calculation step.

[0031]

BEST MODE FOR CARRYING OUT THE INVENTION First, the configuration of a first embodiment of the present invention will be described with reference to FIG.

A document search system to which the present invention is applied includes a display 101 for displaying a search result, a keyboard 102 for inputting a registration / search command and an evaluation for the search result, a central processing unit (CPU) for performing a registration process and a search process. 103, a magnetic disk device 104 for storing various data, a floppy disk drive (FDD) 106 for reading various programs and data from a floppy disk 105, a main memory 107 for temporarily storing programs and data for registration processing and search processing
And a bus 108 connecting them.

The magnetic disk device 104 is one of the secondary storage devices and stores the text 109. FDD10
Information stored in the floppy disk 105 via the main memory 107 or the magnetic disk device 104
Read in. It is also possible to store a program for executing the similarity search method of the present invention in the floppy disk 105, read it, and execute it.

The main memory 107 stores a system control program 110 for controlling the entire search system, a document registration program 111 for registering a document to be searched, and a search control program 112 for controlling the search process.

The search control program 112 is a search condition vector generation program 113 that generates a search condition vector in the first search, a pre-research information storage program 114 that holds the previous search result, and a search condition vector based on the evaluation of the search result. Search condition vector correction program 115 that corrects, a similar search execution program 116 that executes a similar search using the search condition, a search result output program 117 that outputs the result of the similar search, and a search result use end determination that calculates an end determination index. The index calculation program 118 and the end determination index output program 119 that outputs the calculated end determination index are included.

Further, a search condition vector storage area 120 for storing a search condition vector composed of a combination of a characteristic character string used as a search condition for similarity search and its weight, and a pre-research information storage area 121 for storing a previous search result. ,
Also, a work area 122 for storing other temporary data is secured.

In this embodiment, the magnetic disk device is used as the secondary storage device, but another secondary storage device such as a magneto-optical disk device may be used.
A storage device connected via a network such as N (Storage Area Network) may be used.

Further, in this embodiment, the floppy disk 1 is used.
The information stored in 05 is read via the FDD 106, but a reading device corresponding to another storage medium such as a CD-ROM and a CD-ROM drive may be used.

Although a keyboard is used as an input device in this embodiment, a mouse may be used, and a mobile phone or PDA (Personal Digital Assistant) may be used.
Such a mobile terminal may be used, or another input device may be used.

In the present embodiment, the information is output to the display. However, the information may be output to the display of another PC via the network, or the information may be sent as an electronic mail. It does not matter, and a configuration using another output device may be used.

The processing procedure of each program in this embodiment will be described below.

First, the processing procedure of the system control program 110 will be described.

The system control program 110 judges the command input from the keyboard 102, and if it is a document registration command, the document registration program 111.
To start the document registration process. If the command is a document search command, the search control program 112 is activated to execute a document search process.

Next, the processing procedure of the document registration program 111 activated by the system control program 110 will be described.

The document registration program 111, with respect to all the search target documents stored in the floppy disk 105, is an arbitrary identifier (hereinafter, referred to as an identifier that can uniquely identify the search target document in all the search target documents). After being assigned a document identifier), it is stored as the text 109 in the magnetic disk device 104.

Next, the processing procedure of the search control program 112 started by the system control program 110 will be described with reference to the PAD (Problem Analysis Diagram) diagram shown in FIG.

First, in step 401, the search control program 112 is the search condition vector generation program 1
After starting 13 to generate a search condition vector, the search condition vector is stored in the search condition vector storage area 120.

Next, in step 402, the processing from step 403 to step 408 is repeatedly executed until the search end command is input from the keyboard 102.

First, in step 403, the pre-research information storage program 114 is activated, and the search results before the search by the compatibility feedback is performed are stored in the pre-research information storage area 121. Next, the similar search execution program 116 is started, and the search condition vector storage area 1
A similarity search is performed using the search condition vector stored in 20. Next, the search result output program 117 is activated, and the result of the similar search executed by the similar search execution program 116 is output to the display 101. Next, the search result utilization end determination index calculation program 118 is activated to calculate an end determination index serving as a determination index for the end of the compatibility feedback. Next, the end determination index output program 119 is started, and the end determination index calculated by the search result use end determination index calculation program 118 is output to the display 101.

Next, in step 404, the command input from the keyboard 102 is read.

Next, at step 405, the command input at step 404 is judged, and the "desired" or "desired" for the document included in the search result of the similar search output to the display 101 by the search result output program 117. If the command is an "undesired" evaluation command, the search condition vector correction program 115 is started in step 407, and the search condition vector is corrected based on the evaluation command.

In step 406, if the command input in step 404 is judged and the command ends the search, step 408.
In, the contents of the search condition vector storage area 120, the pre-search information storage area 121, and the work area 122 are cleared.

In this embodiment, the evaluation command "desired" or "not desired" is input. However, as disclosed in Japanese Patent Laid-Open No. 2001-22787, the evaluation command "desired" is input. Only the input may be made, or the evaluation may be made by another method such as inputting a desired degree.

Next, the processing procedure of the search condition vector generation program 113 started by the search control program 112 will be described. There is a method disclosed in Related Art 1 as a method for generating a search condition. The processing procedure of the search condition generation program 113 will be described below by taking the search condition generation processing disclosed in Related Art 1 as an example.

The search condition generation program 113 reads a document (hereinafter referred to as seed document) input from the keyboard 102 as a search condition for similarity search into the work area 122, and extracts a characteristic character string from the read seed document. Next, the appearance frequency of the extracted feature character string in the seed document is counted, and the set of the feature character string and the appearance frequency is stored as a search condition vector used as a search condition vector used in the processing of the similar search execution program 116 described later. Area 1
Store in 20.

As a method for extracting the characteristic character string, Japanese Patent Laid-Open No.
The method based on the morphological analysis as disclosed in Japanese Patent Laid-Open No. 301722/1993 may be used.
The method using n-gram as disclosed in 1563 may be used, or another method may be used.

Although the number of types of characteristic character strings extracted from the seed document is not limited in this embodiment, the number of types of characteristic character strings extracted from the seed document is limited to reduce the memory usage. I don't care.

The procedure for generating the search condition vector generated by the search condition vector generation program 113 will be specifically described below with reference to FIG.

First, the seed document 501 is read into the work area 122 by the seed document reading process 502 and stored as seed document data 503.

Next, the characteristic character string extraction / appearance frequency counting process 504 extracts the characteristic character string from the seed document data 503 read by the seed document reading process 502. In this example, a predetermined procedure is used. The characteristic character strings “car”, “accident”, “cellular”, and “phone” are extracted in accordance with the above. After the characteristic character strings are extracted, the appearance frequency of the characteristic character strings in the seed document data 503. Are counted, "car" once, "acciden"
A set 505 of characteristic character strings and appearance frequencies is obtained in the form of “t” once, “cellular” twice, and “phone” three times, and stored in the work area 122.

Finally, by the storage process 506, the set 505 of the characteristic character string and the appearance frequency acquired by the characteristic character string extraction / appearance frequency counting process 504 is obtained as the retrieval condition vector 5.
No. 07 is stored in the search condition vector storage area 120.

Next, the processing procedure of the pre-research information storage program 114 started by the search control program 112 will be described.

First, the pre-research information storage program 114
Clears the contents of the pre-research information storage area 121.
Next, in the processing of the similar search execution program 116, which will be described later, the combination of the rank and the document identifier of the search target document included in the latest search result stored in the work area 122 is searched before performing the re-search by the compatibility feedback. As a result, it is stored in the pre-research information storage area 121.

Although the contents of the pre-research information storage area 121 are cleared in this embodiment, the contents of the pre-research information storage area 121 may not be cleared in order to retain the history. Further, in the present embodiment, the sets of the order and the document identifier of all the documents are stored, but in order to reduce the memory usage, the number of the sets of the order of the stored document and the document identifier is limited. I don't care.

Next, the search condition vector correction program 115 started by the search control program 112 will be described. Search condition vector correction program 115
Modifies the search condition vector stored in the search condition vector storage area 120 based on the user's evaluation. The processing of the search condition vector modification program 115 in the present embodiment can be realized by the method proposed in Prior Art 1 or the like.

Next, the processing procedure of the similar search execution program 116 started by the search control program 112 will be described.

First, the similarity search execution program 116
The degree of similarity between the search condition vector in the search condition vector storage area 120 and the registered document vector generated from all the search target documents stored in the text 109 is expressed by the formula 1
Is calculated using the similarity calculation formula shown in FIG. 1, and the document identifiers of all the search target documents are rearranged in descending order of the similarity, and the work area 1 is used as the search result with the combination of the document identifier and the rank.
It stores in 22.

Although the number of retrieval results to be stored is not limited in this embodiment, the number of retrieval results may be limited in order to reduce the memory usage. Further, in the present embodiment, the combination of the rank and the document identifier is stored as the search result, but other information such as including the similarity in the search result may be added to the search result and stored.

In this embodiment, the set of the order and the document identifier is stored in the work area 122, but it may be temporarily stored in the magnetic disk 104. Further, in the present embodiment, the calculation of the similarity is performed by using the above-mentioned equation 1, but other similarity calculation formulas may be used.

Next, the processing procedure of the search result output program 117 started by the search control program 112 will be described.

The search result output program 117 displays on the display 101 the set of the order of the document to be searched and the document identifier stored in the work area 122 by the similarity search execution program 116. Although the search result is output to the display 101 in this embodiment, the search result may be output to the magnetic disk 104 and used for other processing.

Next, the search result utilization end judgment index calculation program 11 started by the search control program 112
The processing procedure of No. 8 will be described with reference to the PAD diagram shown in FIG.

First, in step 601, the end determination index stored in the work area 122 is reset.

Next, in step 602, the number of documents stored in the pre-retrieval information storage area 121,
The processes of steps 603 and 604 are repeatedly executed.

In step 603, the rank change of unprocessed documents among the documents included in the re-search result stored in the work area 122 is calculated and weighted.
That is, of the documents included in the re-search results stored in the work area 122 by the similarity search execution program 116,
Select one set of unprocessed document rank and document identifier.
Next, the pre-research information storage program 114 refers to the order of the document corresponding to the selected document identifier among the documents included in the pre-research information stored in the pre-research information storage area 121. Next, the rank in the re-search result is subtracted from the rank of the search target document in the search result before the re-search. Next, the calculated difference in rank is divided by the rank of the search target document in the re-search result.

Finally, in step 604, the rank change calculated in step 603 is added to the end determination index.

The end determination index calculated in the process of calculating the end determination index shown in FIG. 6 is expressed by Equation 3. In this formula, M is the number of stored information before re-search, R (D) is the rank of document D in the previous search result, and R '(D) is the rank of document D in the re-search result.

[0078]

[Equation 3] Formula 3 calculates a larger value of the end determination index as the rank variation of the search result is larger, and the end determination index becomes 0 when the rank variation of the search result does not occur.

Although the end determination index is stored in the work area 122 in this embodiment, the end determination index may be stored by another method such as outputting to the magnetic disk 104. In addition, in the present embodiment, the calculation of the termination judgment index is performed by the formula
However, the absolute value of the difference between the ranks may be used, or another calculation formula may be used. Further, in the present embodiment, the end determination index is calculated from the difference in rank, but the end determination index may be calculated from the difference in similarity.

The processing of the retrieval result use end judgment index calculation program 118 will be specifically described below with reference to FIG.

In the example of the case where the document 5 in the fifth place in the search result 301 is evaluated as “desired” and the re-search is performed, the re-search result 302 whose rank greatly changes from the search result 301 is obtained. When the end determination index is calculated using the end determination index calculation formula shown in Formula 3, the end determination index 3.65 is obtained as shown in Formula 4.

[0082]

[Equation 4] Also, in the example in which the re-search from the search result 301 yields the re-search result 303 with a small change in rank,
As shown in, an end judgment index of 0.05 is obtained.

[0083]

[Equation 5] As a result, it is possible to judge the magnitude of the change in rank by referring to the end judgment index.

Next, the processing procedure of the termination judgment index output program 119 started by the search control program 112 will be described. The termination determination index output program 119 displays the termination determination index stored in the work area 122 by the search result utilization termination determination index calculation program 118 on the display 101.

In this embodiment, the termination judgment index is output as a numerical value to the display 101, but the calculated termination judgment index may be displayed as a message 701 as shown in FIG. A message 702 corresponding to the above may be displayed, or may be displayed as a graph 703 of transition of the end determination index from the past. Also, the above information may be output at the same time. Further, in this example, the end judgment index is displayed on the display 101.
Is displayed on the work area 122.
It may be stored in the HDD and used for other processing, or may be stored in the magnetic disk 104 and used for other processing. Further, the user interface for instructing the search may be switched between enabled and disabled according to the end determination index.

The processing flow of this embodiment will be described below with reference to FIG.

First, the seed document 801 is input, and the search condition vector generation program 113 generates a search condition vector 802.

Next, the similar search execution program 116 executes the similar search, and the search result 803 is output.

Next, the user evaluates the search result 803 as “desired” or “not desired”, and the search condition vector correction program 115 causes the search condition vector 802 to be based on the evaluation performed by the user. The search condition vector 802a is modified.

Next, the similar search execution program 116 executes the similar search again, and the re-search result 804 is output.

Next, the search result utilization end judgment index calculation program 118 causes the search result 803 and the re-search result 804.
From this, the end determination index 805 is calculated.

Finally, the end judgment index output program 11
9, the end determination index 805 is output as the message 806.

Although the present embodiment takes the form of document retrieval, the retrieval target may be an image having color information or the like as a characteristic amount, or other electronic data. Further, in the present embodiment, the characteristic character string is extracted from the text at any time, but in a large-scale database, the decrease in search speed becomes a problem. In that case, as disclosed in Japanese Patent Application Laid-Open No. 9-309078, the frequency of appearance of the characteristic character string is stored in advance in the magnetic disk 104 when the document is registered, and the file is referred to in the search. By doing so, high-speed similarity search can be realized. Further, by adding an index to the file, a higher speed search can be realized.

As described above, according to this embodiment, it is possible to easily judge the end of the search by checking the end judgment index, and it is possible to prevent the user from making an incorrect judgment.

Next, the second embodiment will be described.

In the search using the relevance feedback, when the user evaluates a document having a vector similar to the search condition vector among the documents included in the search result,
In some cases, there is little change in the direction of the search condition vector due to evaluation, and there is no change in the ranking in the re-search results. In this case, in the first embodiment, the end determination index is 0, so
There is a problem that the user ends the search even if the target document can be searched by evaluating other documents and continuing the search.

In order to explain the above problem, an example of the process of conformity feedback according to the first embodiment is shown in FIGS. 9 and 10.

In this example, the search condition vector 9 shown in FIG.
It is assumed that the search result 902 is searched by 01. In addition, the user said "Car accident while using a cel
It is assumed that the user wants to search for a document related to "lularphone" and that the document D2 in the search result 902 is evaluated as "desired".

FIG. 10 shows an example of processing when conformity feedback is performed below. Note that, in the search condition vector modification process shown in this example, as disclosed in Prior Art 1, the appearance frequency of the characteristic character string included in the document evaluated as “desired” is used as the weight of the search condition vector. It is supposed to be added.

First, with respect to the search result 902 searched by the similar search execution program 116 based on the search condition vector 901, the user gives an evaluation 1001 that the document D2 is desired.

Next, the retrieval condition vector correction program 1
15, the search condition vector 901 is corrected to the search condition vector 901a based on the evaluation 1001.

Next, the similarity search execution program 116 searches for the re-search result 1002 in which the rank of the document to be searched does not change based on the search condition vector 901a.

Next, the search result use end judgment index calculation program 118 outputs 0 as the end judgment index 1003 as shown in the equation (6).

[0104]

[Equation 6] Therefore, the search condition vector 901a is further modified into the search condition vector 901b by evaluating another document D3 containing similar contents as "desired", and the search result 1006 (searched from the search condition vector 901b ( Hereinafter, even if the rank of the document evaluated as “desired” in the re-retrieval result) increases, the user:
In the evaluation 1004 by the user, the end judgment criterion 100
Since 3 is 0, the end processing 1005 of the compatibility feedback is instructed.

In order to solve the above problems, in the second embodiment, the end judgment index is calculated from the search condition vector. The configuration of the second embodiment of the present invention will be described below with reference to FIG.

In this embodiment, the configuration is almost the same as that of the first embodiment shown in FIG. 1, but the search condition vector use end judgment index calculation program 123 is used in place of the search result use end judgment index calculation program 118. To be Also,
The processing procedure of the pre-research information storage program 114a is different from that of the pre-research information storage program 114.

Next, the pre-retrieval information storage program 114
The processing procedure of a will be described.

First, the pre-retrieval information storage program 114
a clears the contents of the pre-research information storage area 121. Next, in the pre-research information storage area 121, the combination of the characteristic character string and the weight included in the latest search condition vector stored in the search condition vector storage area 120 is set as the search condition vector before the research. Store.

Next, the processing of the retrieval condition vector use end judgment index calculation program 123 will be described.

First, the pre-research information storage program 114
By a, the combination of the characteristic character string and the weight included in the search condition vector before the re-search stored in the pre-research information holding area 121 is read. Next, the search condition vector storage area 120 is searched by the search condition vector correction program 115.
The characteristic character string and weight set included in the search condition vector after re-search stored in are read. Finally, the end determination index is calculated using the end determination index calculation formula shown in Formula 7,
Stored in the work area 122.

[0111]

[Equation 7] Here, w ′ (i) is a weight for the characteristic character string i included in the search condition vector after the re-search, and w ′ (i) is w.
(I) is a weight for the characteristic character string i included in the search condition vector before the re-search, and T1 is the number of different characteristic character strings that match in the search condition vector before and after the re-search (total number of different characteristic character strings). , T2 is the number of different characteristic character strings included in the search condition vector after the re-search, and T3 is the number of different characteristic character strings included in the search condition vector before the re-search.

Equation 7 calculates the inner product of the search condition vector before the re-search and the search condition vector after the re-search, and the direction of the search condition vector is completely changed by the processing by the search condition vector correction program 115. If not, the end determination index becomes 0.

A specific example of this embodiment will be described below with reference to FIG.

In this example, it is assumed that the re-retrieval result 1002 is retrieved in the same manner as the processing flow shown in FIG.
At this point, the search condition vector is (2, 1, 4, 3)
From the search condition vector 901 having the weight of (4, 4,
The search condition vector 901a having the weights of 5 and 3) is modified.

Here, when the search condition vector use end judgment index calculation program 123 calculates the end judgment index based on Expression 7, the end judgment index 1201 becomes 0.078 as shown in Expression 8.

[0116]

[Equation 8] As a result, the user can continue the re-search by referring to the end determination index 1201 even if the rank of the search target document does not change in the re-search result 1002. In this example, the user evaluates the document D3 containing similar contents as “desired” in the evaluation 1202, and the search condition vector correction program 115 causes the search condition vector 901a to further change to the search condition vector 901.
is modified to b, and the similar search execution program 116
By executing the similarity search based on the search condition vector 901b, the re-search result 1006 in which the rank of the document evaluated as "desired" is increased is searched.

In this embodiment, all the characteristic character strings included in the search condition vector used when calculating the end judgment index are used, but the number of characteristic character strings is reduced in order to reduce the memory usage. It does not matter even if the configuration is restricted. Further, in the present embodiment, the formula 7 is used as the termination judgment index calculation formula, but other formulas may be used.

As described above, according to this embodiment, even if there is no change in the rank due to the re-search, the search can be continued without making an erroneous judgment by referring to the end judgment index.

Next, the third embodiment will be described.

In the relevance feedback, when the search is repeated to search for the target document, the weight of the registered document vector evaluated as desired remains unchanged even if the search is repeated. Will increase.
Therefore, as the search is repeated, the difference between the weight of the registered document vector and the weight of the search condition vector increases.

As a result, when the difference becomes large, the state of the search condition vector does not change greatly even if the conformity feedback is performed, and the search result does not change (hereinafter, referred to as a converged state).

When the convergence state is reached, the user should end the conformity feedback, and if the search result is dissatisfied, it is necessary to re-execute the conformance feedback and further consider the search method itself.

In the second embodiment, if there is no change in the search condition vector, the search should be ended because the convergence state has been reached, or the search target document having a vector similar to the search condition vector is evaluated. Therefore, there is a problem that other documents cannot be evaluated to determine whether the search should be continued, and the end of the search is erroneously determined.

In order to solve the above problem, in this embodiment,
The end determination index calculated after performing the re-search is calculated in advance before the re-search is executed and presented to the user.

First, the structure of this embodiment will be described with reference to FIG.

This embodiment has almost the same configuration as that of the first embodiment shown in FIG. 1, except that the prediction end judgment index calculation program 124 is used instead of the search result use end judgment index calculation program 118. Further, a search condition vector correction prediction program 125 is newly added. In addition, a prediction search condition vector storage area 126 is newly secured. The processing procedure of the end determination index output program 119a is different from that of the end determination index output program 119.

Next, the prediction end judgment index calculation program 1
The processing procedure of 24 will be described using the PAD diagram shown in FIG.

First, in step 1401, a combination of evaluation commands that can be specified (hereinafter referred to as an evaluation pattern) is acquired from the set of the order of the document to be searched and the document identifier stored in the work area 122 by the similar search execution program 116. Then, it is stored in the work area 122.

A specific example of evaluation pattern acquisition will be described below with reference to FIG.
This will be described using 5.

In this example, it is assumed that two documents are included in the search result 1501. When two kinds of evaluation commands “desired” and “not desired” can be designated for the two documents in the search result, there are a total of eight possible evaluation patterns 1502 including the case where no command is designated. Become.

Next, at step 1402, as many as the number of evaluation patterns stored in the work area 122,
The processing from step 1403 to step 1406 is repeatedly executed.

At step 1403, one unprocessed evaluation pattern is selected.

Next, in step 1404, the contents of the search condition vector storage area 120 modified by the search condition vector modification program 115 are copied to the predicted search condition vector storage area 126.

Next, in step 1405, the retrieval condition vector correction prediction program 125 is activated to retrieve the retrieval condition vector in the predicted retrieval condition vector storage area 126 by re-retrieval when the retrieval pattern is designated. Correct the condition vector (hereinafter referred to as the predictive search condition vector).

Finally, in step 1406, the equation 7
In the search condition vector storage area 126, the search condition vector w '(i) after the re-search in the search condition vector w (i) in the search condition vector storage area 126 By using the search condition vector, the end determination index (hereinafter, referred to as a predicted end determination index) when the re-search is performed based on the evaluation command is calculated based on Formula 7, and stored in the work area 122.

In this embodiment, all combinations of evaluation commands for search results are acquired. However, in order to reduce the memory usage, the number of search target documents that can be evaluated is limited and the combination is performed. The number may be limited.

Next, the prediction end judgment index calculation program 1
The processing of the search condition vector correction prediction program 125 started by 24 will be described.

Search condition vector correction prediction program 12
5 shows the search condition vector stored in the predicted search condition vector storage area 126 based on the evaluation pattern selected by referring to the predicted end determination index in the process of the predicted end determination index calculation program 124 shown in FIG. Similarly, modify the search condition vector.

Next, the end judgment index output program 119
The processing of a will be described.

The end judgment index output program 119a displays the prediction end judgment index stored in the work area 122 by the prediction end judgment index calculation program 124, for example, as shown in FIG.
As shown in, a display 150 is displayed as a table 1503 of the prediction end determination index for each evaluation pattern.

In this embodiment, the prediction end judgment index is calculated from the search condition vector, but the prediction end judgment index may be calculated from the rank of the document to be searched, or the prediction end judgment index may be calculated from other information. You may calculate. Further, although the prediction end determination index is output for each evaluation pattern in the present embodiment, the prediction end determination index may be output in another format such as an average output. Further, in the present embodiment, the prediction end determination index for all the evaluation patterns is output, but the prediction end determination index may be output at the timing when the evaluation command is input.

According to this embodiment, it is possible to predict the change of the search condition vector due to the re-search before executing the re-search. Therefore, even if there is no change in the search result and the search condition vector, another document is evaluated. By doing so, you can know that the re-search can be continued, and you can continue the re-search without making a wrong decision.

[0143]

By referring to the end determination index presented by the system, it is possible to efficiently determine the end of the search in the compatibility feedback, and the user can perform the search with a small load.

[Brief description of drawings]

FIG. 1 is a diagram showing an overall configuration of a first embodiment of a document search system according to the present invention.

FIG. 2 is a diagram showing an outline of similar document search and conformity feedback processing according to the related art 1.

FIG. 3 is a diagram showing an example of a case in which the rank variation due to the compatibility feedback is large and a case in which the rank variation is small.

FIG. 4 is a PAD showing a processing procedure of a search control program according to the first embodiment of the present invention.

FIG. 5 is a diagram showing a procedure for generating a search condition vector in the first embodiment of the present invention.

FIG. 6 is a PAD showing a processing procedure of a search result use end judgment index calculation program in the first embodiment of the present invention.

FIG. 7 is a diagram showing a display example of an end determination index in the first embodiment of the present invention.

FIG. 8 is a diagram showing a flow of processing in the first embodiment of the present invention.

FIG. 9 is a diagram showing an example of similar document search according to the second exemplary embodiment of the present invention.

[Fig. 10] Fig. 10 is a diagram illustrating an example of processing of conformance feedback according to the first embodiment in the second embodiment of the present invention.

FIG. 11 is a diagram showing a configuration of a search subsystem according to a second embodiment of the present invention.

FIG. 12 is a diagram showing an example of processing of conformity feedback in the second exemplary embodiment of the present invention.

FIG. 13 is a diagram showing a configuration of a search subsystem according to a third embodiment of the present invention.

FIG. 14 is a PAD showing a processing procedure of a prediction end determination index calculation program according to the third embodiment of the present invention.

FIG. 15 is a diagram showing a display procedure of a prediction end determination index in the third embodiment of the present invention.

[Explanation of symbols]

101 Display 102 Keyboard 103 Central Processing Unit (CPU) 104 Magnetic Disk Unit 105 Floppy Disk 106 Floppy Disk Drive (FDD) 107 Main Memory 108 Bus 110 System Control Program 111 Document Registration Program 112 Search Control Program 113 Search Condition Vector Generation Program 114 Pre-research information storage program 115 Search condition vector correction program 116 Similar search execution program 117 Search result output program 118 Search result use end judgment index calculation program 119 End judgment index output program 120 Search condition vector storage area 121 Pre-search information storage area 122 work area 123 search condition vector use end judgment index calculation program 124 prediction end judgment index calculation program 125 Search condition vector correction prediction program 126 Prediction search condition vector storage area

   ─────────────────────────────────────────────────── ─── Continued front page    (72) Inventor Tadataka Matsubayashi             890 Kashimada, Sachi-ku, Kawasaki City, Kanagawa Stock             Hitachi, Ltd. Business Solutions             Within the department (72) Inventor Yasuhiko Inaba             890 Kashimada, Sachi-ku, Kawasaki City, Kanagawa Stock             Hitachi, Ltd. Business Solutions             Within the department (72) Inventor Yuichi Ogawa             890 Kashimada, Sachi-ku, Kawasaki City, Kanagawa Stock             Hitachi, Ltd. Business Solutions             Within the department (72) Inventor Shinya Yamamoto             5030 Totsuka Town, Totsuka Ward, Yokohama City, Kanagawa Prefecture             Ceremony company Hitachi Ltd. software division (72) Inventor Masayuki Hamakawa             Tokyo Stock Exchange, 1-6-27 Shinsuna, Koto-ku, Tokyo             Hitachi, Ltd. Public Systems Division F-term (reference) 5B075 PR03 QM08

Claims (11)

[Claims] 1. A similarity search method using relevance feedback searches a database using a search condition vector consisting of a feature quantity and a weight of the feature quantity as a search condition, and searches the database for the search results obtained by the search. Receiving at least one evaluation of “desired” or “undesired” input by the user, changing the weight of the feature amount of the search condition vector based on the evaluation, and performing a re-search. Similarity search characterized by storing information after re-search and calculating an end judgment index as a judgment index for ending the search by comparing the stored information before and after the re-search by a predetermined method Method. 2. A similarity search method using relevance feedback searches a database using a search condition vector consisting of a set of feature quantities and weights of the feature quantities, and a user searches the search results obtained by the search. Receive at least one of the input "desired" or "not desired", change the weight of the feature amount of the search condition vector based on the evaluation, perform a re-search, and store the information before the re-search, The information after the re-search is predicted and stored based on the stored information before the re-search, and the information is searched by comparing the stored information before the re-search with the predicted information after the re-search by a predetermined method. A similarity search method, characterized in that an end judgment index serving as a judgment index for terminating is calculated. 3. The similarity search method according to claim 1, wherein the search condition vector is a feature character string that may be an independent word included in a document given as a search condition, and the appearance of the feature character string. A similarity search method, characterized in that the appearance information of a characteristic character string included in the search condition vector is changed based on the evaluation, and a search is performed again. 4. The similarity search method according to claim 1, 2 or 3, wherein the information stored before and after the search is
A similarity search method, characterized in that at least one of data rank and data similarity included in a search result and a set of identifiers that can uniquely identify data in a database. 5. The similarity search method according to claim 1, wherein the information to be stored before and after the search is a search condition vector. 6. The similarity search method according to claim 1, 2, or 3, characterized in that information for prompting continuation or termination of the search based on the calculated termination judgment index is displayed via a user interface. A similar search method. 7. The similarity search method according to claim 1, 2 or 3, wherein whether or not to display a user interface for instructing a search is controlled based on the calculated end determination index. 8. A computer-readable portable storage medium that stores a program for executing a similarity search method using compatibility feedback, the similarity search method comprising:
Pre-research information storage step of storing information before re-search, post-research information storage step of storing information after re-search, and information before and after re-search stored in the above step by a predetermined method. A computer-readable portable storage medium comprising an end determination index calculating step of calculating an end determination index serving as a determination index for ending a search by comparing. 9. A computer-readable portable storage medium that stores a program for executing a similarity search method using compatibility feedback, the similarity search method comprising:
Pre-research information storage step of storing information before re-search, pre-research information prediction step of acquiring and storing post-research information before re-search execution by a user, and before re-search stored in the above step Computer-readable, characterized in that it has an end judgment index calculating step for calculating an end judgment index serving as a judgment index for terminating the search by comparing the above information with the predicted information after the re-search by a predetermined method. Portable storage medium. 10. A similarity search device based on relevance feedback, a search means for searching a database using a search condition vector, “desired” or “not desired” input by a user with respect to a search result obtained by the search. "
Evaluation receiving means for receiving at least one evaluation of,
Re-search means for performing re-search by changing the weight of the feature quantity of the search condition vector based on the evaluation, pre-re-search information storage means for storing information before re-search, re-search for storing information after re-search After-information storage means, Completion judgment for calculating an end-judgment index which is a judgment index for terminating the search by comparing the information before and after the re-search stored before and after the re-search with a predetermined method A similarity search device having index calculation means. 11. A similarity search device using relevance feedback, a search means for searching a database using a search condition vector, “desired” or “not desired” input by a user with respect to a search result obtained by the search. "
Evaluation receiving means for receiving at least one evaluation of,
Re-search means for changing the weight of the feature quantity of the search condition vector based on the above-mentioned evaluation to perform re-search, pre-re-search information storage means for storing information before re-search, and re-search for information after re-search by user Pre-research information prediction means to be acquired and stored before execution, pre-research information acquired by the pre-search information storage means and pre-research prediction information acquired by the post-research information prediction means are determined in advance. A similarity search device, comprising: an end determination index calculating means for calculating an end determination index serving as a determination index for ending the search by performing a comparison by the above method.